chronODE : A framework to integrate time-series multi-omics data based on ordinary differential equations combined with machine learning

Beatrice Borsari, Mor Frank, Eve S. Wattenberg, Ke Xu, Susanna Liu, X. Yu, Mark GersteinbioRxiv 2023 · DOI

TL;DR

This paper introduces chronODE, a novel framework that uses ordinary differential equations (ODEs) and machine learning to model the dynamic changes in gene expression and chromatin accessibility over time. It successfully predicts gene expression from chromatin changes and reveals how regulatory element kinetics are tied to specific cell types during brain development.

What Problem Does This Paper Address?

Most functional genomics studies capture a snapshot of molecular processes, limiting our understanding of dynamic events like cell differentiation. While longitudinal studies offer more insight, modeling the complex, time-dependent kinetics of gene expression and epigenetic changes, especially across different data types (multi-omics), remains a significant challenge. Existing methods often rely on "latent time" derived from single-cell snapshots, which can limit the biological interpretability and accuracy of kinetic parameter estimation. There's a need for a framework that can directly model real time-series data, integrate multiple omics layers, and provide biologically meaningful kinetic parameters for regulatory elements and genes.

Methods

The core of this work is the chronODE framework, which leverages ordinary differential equations (ODEs) to model the rate of change in chromatin accessibility and gene expression over time.

  • ODE Modeling: chronODE uses a logistic curve-based ODE to model these dynamic changes. This approach is flexible enough to capture various patterns, including log-like and exponential trends, and provides biologically interpretable parameters like the rate of change (k) and asymptote (b).
  • Time-Series Interpolation: To handle sparse time-series data, the framework employs linear interpolation to generate a denser time course before fitting the ODEs.
  • Neural Network for Epigenome-Transcriptome Linkage: A neural network architecture is used to learn the complex, non-linear relationships between time-series derivatives of chromatin accessibility and gene expression. This allows for predicting gene expression changes based on chromatin accessibility dynamics.
  • Data Integration: The framework integrates DNase-seq and ATAC-seq data for chromatin accessibility, and RNA-seq data for gene expression, using joint normalization and batch correction techniques.

Datasets

The study utilized the following datasets:

  • Mouse Brain Development Time-Course Data:
    • DNase-seq and ATAC-seq data for chromatin accessibility.
  • PolyA+ RNA-seq data for gene expression.
  • These data were generated across eight developmental time points (E10.5 to postnatal day 1) in three mouse brain regions: forebrain, midbrain, and hindbrain.
  • Single-Cell (sc) ATAC-seq Data:
    • Forebrain scATAC-seq data from the same developmental time course was used to identify cell-type specific regulatory elements.

Key Insights and Ideas

  • Region-Specific Timing of Epigenome Rewiring: Chromatin accessibility changes during brain development are not uniform across different brain regions, suggesting distinct temporal programs for forebrain, midbrain, and hindbrain.
  • Cell-Type Specific Kinetics: The timing and nature of chromatin accessibility changes (increasing vs. decreasing) are strongly correlated with the emergence of specific cell types. Early-appearing cell types (like radial glia) are associated with early-acting regulatory elements, while later-emerging cell types (like mature neurons) are linked to regulatory elements with later activation kinetics.
  • Predictive Power of Chromatin Dynamics: Changes in chromatin accessibility at regulatory elements can accurately predict the rate of change in target gene expression over time, highlighting a causal link.
  • Multi-Pattern cCREs Drive Brain Development: Genes regulated by multiple types of dynamic cis-regulatory elements (cCREs) are enriched in brain-specific functions and are more dynamic during development, suggesting a complex regulatory network for key developmental processes.
  • Activator vs. Repressor Roles: The framework can distinguish between cCREs that positively correlate with gene expression (likely involving activators) and those that negatively correlate (likely involving repressors), providing insights into regulatory mechanisms.

Key Results

  • Identification of Dynamic cCREs: Out of over 926,000 cCREs, 405,554 were identified as "active" and 24% showed dynamic changes in chromatin accessibility across the three brain regions.
  • Kinetic Pattern Classification: chronODE successfully modeled the time-series trajectories of over 99% of dynamic cCREs, classifying them into logistic, log-like, or exponential patterns.
  • Predictive Accuracy: The neural network model achieved a correlation of 0.71 ± 0.29 for positively correlated cCRE-gene pairs and 0.57 ± 0.29 for negatively correlated pairs in predicting gene expression derivatives. This outperformed a Random Forest model.
  • Cell-Type Specificity: Analysis of scATAC-seq data revealed that increasing cCREs specific to late-emerging cell types (e.g., eIN3, eAC) showed a higher proportion of late (exponential) trajectories, while ubiquitous cCREs were enriched for early (log-like) trajectories.
  • Gene Function Enrichment: Genes linked to multi-pattern cCREs were significantly enriched for neurogenesis, trans-synaptic signaling, and sensory system development functions.

Limitations and Future Work

  • Single cCRE Scenario: The current predictive model primarily focuses on a single cCRE regulating a single gene. The authors anticipate extending this to more complex scenarios involving multiple cCREs or chromatin features.
  • Bulk vs. Single-Cell: While the framework successfully infers kinetics from bulk data, integrating time-resolved single-cell measurements is a promising future direction.
  • Broader Applications: The authors envision applying chronODE to various biological systems and data modalities to understand molecular pathways, identify drug targets, and study mechanisms underlying drug resistance.
  • Therapeutic Strategies: The long-term goal is to leverage these kinetic maps of transcriptional and epigenetic processes to design tailored therapeutic strategies.

Summary updated

Abstract

Abstract Most functional genomic studies are conducted in steady-state conditions, therefore providing a description of molecular processes at a particular moment of cell differentiation or organismal development. Longitudinal studies can offer a deeper understanding of the kinetics underlying epigenetic events and their contribution to defining cell-type-specific transcriptional programs. Here we develop chronODE , a mathematical framework based on ordinary differential equations that uniformly models the kinetics of temporal changes in gene expression and chromatin features. chronODE employs biologically interpretable parameters that capture tissue-specific kinetics of genes and regulatory elements. We further integrate this framework with a neural-network architecture that can link and predict changes across different data modalities by solving multivariate time-series regressions. Next, we apply this framework to investigate region-specific kinetics of epigenome rewiring in the developing mouse brain, and we demonstrate that changes in chromatin accessibility within regulatory elements can accurately predict changes in the expression of putative target genes over the same time period. Finally, by integrating single-cell ATAC-seq data generated during the same time course, we show that regulatory elements characterized by fast activation kinetics in bulk measurements are active in early-appearing cell types, such as radial glial and other neural progenitors, whereas elements characterized by slow activation kinetics are specific to more differentiated cell types that emerge at later stages of brain development.
BibTeX
@misc{borsari2023i,
  author = {Beatrice Borsari and Mor Frank and Eve S. Wattenberg and Ke Xu and Susanna Liu and X. Yu and Mark Gerstein},
  title = {{<i>chronODE</i> : A framework to integrate time-series multi-omics data based on ordinary differential equations combined with machine learning}},
  year = {2023},
  doi = {10.1101/2023.12.13.571513},
  howpublished = {bioRxiv},
}

Loading related content...

Figure